An ai-driven multi-path hybrid supply-demand matching sorting method and system

Abstract

This invention relates to the field of supply and demand matching technology, specifically to an AI-driven multi-path hybrid supply and demand matching ranking method and system. The method includes the following steps: receiving and analyzing natural language demand text to generate structured demand elements; constructing multi-path query expressions when the structured demand elements meet preset completeness conditions; generating demand clarification information and updating it according to supplementary input when the conditions are not met; performing hybrid recall in the recall path based on the multi-path query expressions; fusing dense vector retrieval results and sparse vector retrieval results and re-ranking them; performing inverse ranking fusion and product subject screening on the re-ranked candidate set within the path; inputting the screened candidate objects and structured demand elements into an LLM model for ranking; and outputting the ranking results of matching enterprises. This invention, through demand structuring and multi-path hybrid recall, makes the candidate results form the basis for ranking and makes the results subject to the demand subject and constraint attributes.

Description

Technical Field

[0001] This invention relates to the field of supply and demand matching technology, and in particular to an AI-driven multi-path hybrid supply and demand matching and sorting method and system. Background Technology

[0002] The technology of supply and demand matching involves object modeling, condition expression, candidate retrieval, result filtering, and ranking output between demanders and suppliers. It is widely used in scenarios such as enterprise procurement, supply chain management, industrial collaboration, project undertaking, technology transfer, investment promotion services, and public service matchmaking. Among them, the traditional multi-path hybrid supply and demand matching ranking method refers to receiving product, service, or project requirements input by demanders in the supply and demand matching platform, transforming the demand subject, keywords, geographical conditions, performance parameters, and business constraints into structured search conditions, and then using keyword retrieval, tag filtering, semantic retrieval, and other means in enterprise information databases, product information databases, capability information databases, and attribute information databases to uniformly rank the candidate suppliers, product objects, or service objects obtained from multiple recall paths to form a candidate list.

[0003] Traditional supply and demand matching platforms rely on keyword retrieval, tag filtering, and semantic retrieval to process demand text. The demand subject, search clues, and constraint attributes are easily mixed up. Enterprise information, product information, and attribute information are scattered in different fields. The candidate results formed by different recall paths lack a unified ranking basis. Candidates with similar semantics but different product subjects are easy to enter the ranking results. Constraints such as region, performance, and parameters are difficult to be integrated into the recall and ranking process. The candidate selection basis is unclear, resulting in a deviation between the ranking results of matched enterprises and the actual demand. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-driven multi-path hybrid supply and demand matching and ranking method and system, which is used to perform structured processing of natural language demand text in supply and demand matching scenarios. This allows the demand subject, retrieval clues, and constraint attributes to participate in query expression construction, hybrid recall, ranking fusion, and subject screening processes, respectively. The system also uses an LLM model to rank the selected candidate objects, thereby providing a generation method with clear data sources and processing links for the ranking results of matching enterprises, thus solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an AI-driven multi-path hybrid supply and demand matching and sorting method, comprising the following steps: S1: Receive the natural language requirement text input by the user, perform requirement analysis on the natural language requirement text, generate structured requirement elements that represent the requirement subject, retrieval clues and constraint attributes, and enter the query expression construction when the structured requirement elements meet the preset completeness conditions. When the structured requirement elements do not meet the preset completeness conditions, generate requirement clarification information and update the structured requirement elements according to the supplementary input. S2: Based on the structured demand elements, construct a multi-way query expression according to the supply-side profile field, product entity field, and constraint attribute field; S3: Based on the multi-way query expression, perform hybrid recall in the corresponding recall path. In each recall path, merge the dense vector retrieval results and the sparse vector retrieval results, and reorder the fused candidate results to form a reordered candidate set within the path. S4: Merge the candidate sets for reordering within each recall path, perform reciprocal sorting and fusion based on the candidate object's ranking position in the corresponding recall path, obtain a preliminary fusion ranking result, filter the candidate objects for product subjects based on the subject association information and product standardized name in the structured demand elements, input the candidate objects after product subject filtering and the structured demand elements into the LLM model for ranking, and output the ranking result of matching enterprises.

[0006] Furthermore, the processing procedure for the requirement analysis is as follows: the natural language requirement text is divided into text segments; candidate elements are extracted according to requirement object terms, retrieval expression terms, and constraint description terms; the candidate elements are written into requirement object terms, retrieval clue terms, and constraint attribute terms; and candidate elements with missing fields, field conflicts, or that cannot be mapped to the recall path are marked as elements to be clarified; requirement clarification information is generated based on the elements to be clarified; the structured requirement elements are updated after receiving the supplementary input; and the updated structured requirement elements are used as the data basis for determining whether the preset completeness conditions are met.

[0007] Further, the process of determining whether the structured requirement element meets the preset completeness condition is as follows: according to the field mapping relationship corresponding to the recall path, the requirement object item, the search clue item, and the constraint attribute item are mapped and verified. When the requirement object item, the search clue item, and the constraint attribute item can all be mapped to at least one recall path and there is no field conflict, it is determined that the structured requirement element meets the preset completeness condition. When any element item cannot be mapped to the recall path or there is a field conflict, the corresponding element item is determined as the element to be clarified.

[0008] Furthermore, the process of constructing the multi-way query expression is as follows: based on the structured demand elements, a field mapping relationship is established between element items and recall fields; according to the field mapping relationship, element items representing supply capacity are mapped to the supply-side profile field, element items representing product entities are mapped to the product entity field, and element items representing business constraints are mapped to the constraint attribute field; based on the field mapping results, the multi-way query expression with recall field identifiers is generated respectively, and the multi-way query expression is assigned to the recall path corresponding to the recall field identifier.

[0009] Further, the process of performing the hybrid recall is as follows: in each recall path, a query expression corresponding to the recall path is selected from the multi-way query expression, and a dense vector retrieval request and a sparse vector retrieval request are generated respectively. Dense vector retrieval results are obtained according to the dense vector retrieval request, and sparse vector retrieval results are obtained according to the sparse vector retrieval request. The dense vector retrieval results and the sparse vector retrieval results are merged into a fused candidate result according to the fusion weight corresponding to the recall path, and the fused candidate result is reordered.

[0010] Furthermore, the process of forming the candidate set for in-path reordering is as follows: extract the candidate object text corresponding to each candidate object in the fused candidate results, and combine the natural language requirement text and the candidate object text to form an association determination input. By calculating the matching degree of the association determination input, the in-path association ranking of each candidate object in the corresponding recall path is obtained. When the same candidate object appears repeatedly in the same recall path, the candidate object record with the higher in-path association ranking is retained, and the candidate set for in-path reordering is formed based on the retained candidate object record.

[0011] Furthermore, the process of performing inverse sorting fusion on the candidate sets of the path reordering output by each recall path is as follows: the candidate object records in different recall paths are merged according to the candidate object identifier to obtain a cross-path candidate object set; the path ordering position of each candidate object in the cross-path candidate object set in each recall path is read; the candidate objects that do not appear in the corresponding recall path are marked as non-participating; a fusion sorting score is generated according to the number of recall paths in which the candidate object is in the participating state, the path ordering position, and the recall path weight; and the preliminary fusion sorting result is formed according to the fusion sorting score.

[0012] Further, the process of filtering product subjects based on the preliminary fusion and sorting results is as follows: the candidate object records are read according to the sorting order in the preliminary fusion and sorting results, and the product name field, product type field, and product standardized name are extracted from the candidate object records. The consistency of the subject association information in the structured requirement elements with the product standardized name is compared to obtain the subject consistency result. When the subject consistency result meets the preset consistency condition, the corresponding candidate object is retained. When the subject consistency result does not meet the preset consistency condition, the corresponding candidate object is downgraded, and a filtered candidate object set is formed based on the retained or downgraded candidate objects.

[0013] Furthermore, the process of inputting the filtered candidate object set and the structured demand elements into the LLM model for sorting is as follows: the demand subject, the search clues, and the constraint attributes in the structured demand elements are assembled with the supply-side profile information, product entity information, and constraint attribute matching information of each candidate object in the filtered candidate object set to form a candidate sorting input. The LLM model performs object-by-object sorting judgment on the candidate sorting input to obtain the sorting score corresponding to each candidate object. The candidate objects are then arranged in descending order according to the sorting score, and the sorted candidate objects are determined as the matching enterprise sorting result.

[0014] Furthermore, an AI-driven multi-path hybrid supply and demand matching and sorting system is provided. This system is applied to the AI-driven multi-path hybrid supply and demand matching and sorting method. The system includes a demand processing module, a recall processing module, a sorting and filtering module, and a fine sorting output module connected along the data flow direction.

[0015] The demand processing module receives natural language demand text input by the user, performs demand analysis on the natural language demand text, generates structured demand elements representing the demand subject, search clues, and constraint attributes, and generates demand clarification information when the structured demand elements do not meet preset completeness conditions. The recall processing module constructs a multi-way query expression based on the structured demand elements, according to the supply-side profile field, product entity field, and constraint attribute field. The ranking and filtering module performs hybrid recall in the corresponding recall path based on the multi-way query expression and forms a path-in-path re-ranking candidate set. The fine-ranking output module merges, performs inverse sorting fusion, product subject filtering, and LLM model ranking on the path-in-path re-ranking candidate sets output by each recall path, and outputs the matching enterprise ranking results.

[0016] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: By transforming natural language requirement text into structured requirement elements and triggering query expression construction or requirement clarification with preset completeness conditions, the input before recall can form a searchable data foundation; by driving hybrid recall of corresponding recall paths through multi-way query expressions, dense vector retrieval results and sparse vector retrieval results are first merged and reordered within the path, and then merged in reverse order according to the sorting position, so that candidate results from different recall paths can enter the same sorting chain; through product subject screening and LLM model sorting, the matching enterprise sorting results can be simultaneously constrained by subject association information, product standardized name and structured requirement elements. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram illustrating the requirements structuring and requirements clarification processes of this invention; Figure 3 This is a multi-path heterogeneous hybrid recall data flow diagram of the present invention; Figure 4 This is a schematic diagram of the RRF fusion screening and LLM model fine ranking of the present invention; Figure 5 This is a system module collaboration diagram of the present invention. Detailed Implementation

[0019] The technical solution described in this invention will be further explained below with reference to the accompanying drawings. This invention is applicable to scenarios such as industrial supply and demand matching, enterprise procurement and supply chain management, project collaboration, technology transfer, investment promotion services, and public service matchmaking. The processing objects are natural language demand text input by the demand side and supply-side enterprise information, product information, capability information, and attribute information. In the following description, natural language demand text refers to product, service, or project demands input by users in natural language; multi-way query expression refers to retrieval expressions constructed for different supply-side fields or recall targets; recall paths include enterprise profile recall paths, product name recall paths, product original description recall paths, and product attribute recall paths; candidate objects refer to candidate enterprises, candidate products, or supply objects obtained through the recall paths; Top-N matching enterprise results refer to the ranking results output after preliminary sorting and screening and LLM model fine-tuning.

[0020] In this embodiment, the Milvus vector database is used to store and retrieve supply-side vector data, and the BGE-M3 model is used to convert demand text and supply-side text into dense and sparse vector representations. Dense vector representations characterize the semantic relevance between demand and candidate objects, while sparse vector representations characterize the explicit hit relationships of keywords, terms, and entity names. Reranker is used to perform a secondary ranking of the relevance between candidate objects and demand text after initial recall. RRF stands for Reverse Rank Fusion, used to fuse multi-path results based on the ranking position of candidate objects in each recall path.

[0021] Please see Figures 1 to 4 This embodiment provides an AI-driven multi-path hybrid supply and demand matching and ranking method. This method is executed by the supply and demand matching platform and forms a Top-N matching enterprise result in the order of demand structuring, query construction, multi-path heterogeneous hybrid recall, preliminary ranking and screening, and LLM model fine ranking.

[0022] S1: Receives natural language requirement text input by the user, performs requirement analysis on the natural language requirement text, generates structured requirement elements that represent the requirement subject, search clues and constraint attributes, enters query expression construction when the structured requirement elements meet the preset completeness conditions, generates requirement clarification information when the structured requirement elements do not meet the preset completeness conditions, and updates the structured requirement elements according to supplementary input.

[0023] In S1, users input natural language requirement text into the input interface of the supply and demand matching platform. This text can include product name, target audience, performance requirements, geographical scope, configuration requirements, parameter requirements, or a description of business capabilities. After receiving the natural language requirement text, the supply and demand matching platform performs requirement analysis. The analysis first divides the natural language requirement text into text segments and extracts candidate elements based on requirement object terms, search expression terms, and constraint description terms.

[0024] The demand object term is used to identify the core product or service object to which the demand is directed and is written into the demand object term; the search expression term is used to retain keywords, industry terms, aliases or standard names explicitly expressed by the user and is written into the search clue term; the constraint description term is used to characterize dynamic business conditions such as company capabilities, product configuration, technical parameters, geographical conditions or business scope and is written into the constraint attribute term.

[0025] After completing the basic analysis, the supply and demand matching platform performs a completeness check on the structured demand elements. This check involves verifying the mapping of demand object items, search clue items, and constraint attribute items according to the field mapping relationships corresponding to the recall paths. When all demand object items, search clue items, and constraint attribute items can be mapped to at least one recall path, and there are no missing fields, field conflicts, or unenforceable constraints, the structured demand elements are determined to meet the preset completeness conditions, and the process proceeds to query expression construction.

[0026] When natural language demand text contains incomplete information, unclear core objects, missing regions or core parameters, significant semantic ambiguity, or unenforceable constraints, the supply-demand matching platform marks the corresponding candidate elements as elements requiring clarification and generates demand clarification information based on these elements. This clarification information guides users to supplement missing core information or eliminate ambiguity. After receiving supplementary input, the supply-demand matching platform writes it back to the corresponding demand object item, search clue item, or constraint attribute item, and uses the updated structured demand elements as the data basis for constructing subsequent query expressions.

[0027] S2: Based on the structured demand elements, construct a multi-way query expression according to the supply-side profile field, product entity field, and constraint attribute field.

[0028] In S2, the supply and demand matching platform establishes a field mapping relationship between element items and recall fields based on structured demand elements. The supply-side profile field carries supply-side profile information such as enterprise description, enterprise capabilities, and business scope; the product entity field carries product object information such as product name, product alias, product standardized name, and product description; and the constraint attribute field carries constraint information such as product attributes, technical characteristics, capability characteristics, geographical scope, configuration conditions, or parameter conditions.

[0029] Based on field mapping relationships, the supply and demand matching platform constructs query expressions for different recall targets according to the same structured demand elements. For the enterprise profile recall path, query expressions biased towards enterprise capabilities and business scope are generated to discover enterprise objects whose overall capabilities are close to the semantics of the demand. For the product name recall path, query expressions biased towards product entity names, product aliases, and product standardized names are generated to capture the direct correspondence between demand subjects and supply product entities. For the original product description recall path, query expressions biased towards the original product description text, usage description, and scenario description are generated to utilize fine-grained information in the original data from the supply side. For the product attribute recall path, query expressions biased towards technical parameters, configuration conditions, attribute characteristics, and capability characteristics are generated to enhance the participation of dynamic attribute constraints in the recall stage.

[0030] Each query expression is marked with a recall field identifier. The recall field identifier indicates the recall path that the query expression should enter. After multiple query expressions are generated, the supply and demand matching platform assigns different query expressions to the enterprise profile recall path, product name recall path, product original description recall path, and product attribute recall path, respectively, so that the same natural language requirement can enter the candidate discovery process from different perspectives such as enterprise capabilities, product entities, product descriptions, and attribute conditions.

[0031] S3: Based on the multi-way query expression, perform hybrid recall in the corresponding recall path. In each recall path, the dense vector retrieval results and the sparse vector retrieval results are fused, and the fused candidate results are reordered to form a reordered candidate set within the path.

[0032] In S3, the supply and demand matching platform uses the Milvus vector database as the foundation for storing and retrieving supply-side vector data. After data such as supply-side enterprise descriptions, enterprise capabilities, business scope, product names, product aliases, standardized product names, original product descriptions, product attributes, technical characteristics, and capability characteristics are processed into vector retrieval data that can be used by various recall paths.

[0033] The supply and demand matching platform uses the BGE-M3 model to vectorize query expressions and supply-side text. The BGE-M3 model can output both dense and sparse vector representations simultaneously. Dense vector retrieval is used to measure the deep semantic similarity between demand and candidate objects; sparse vector retrieval is used to characterize explicit hit relationships of keywords, terms, product names, product aliases, and entity names.

[0034] In each recall path, the supply-demand matching platform selects the query expression corresponding to that recall path from multiple query expressions, generating dense vector retrieval requests and sparse vector retrieval requests respectively. Based on the dense vector retrieval request, dense vector retrieval results are obtained from the Milvus vector database; based on the sparse vector retrieval request, sparse vector retrieval results are obtained. Subsequently, the supply-demand matching platform merges the dense vector retrieval results and the sparse vector retrieval results into a fused candidate result according to the fusion weight corresponding to the recall path. The fusion weight is used to determine the influence direction of semantically similar results and explicit hit results within the same path.

[0035] The merged candidate results are then re-ranked by Reranker. Reranker uses natural language demand text and candidate object text as inputs for relevance determination. The candidate object text originates from candidate object records in the corresponding recall paths. In the enterprise profile recall path, this corresponds to enterprise description, enterprise capabilities, and business scope; in the product name recall path, it corresponds to product name, product alias, and standardized product name; in the product original description recall path, it corresponds to the original product description text; and in the product attribute recall path, it corresponds to product attributes, technical features, or capability features. Reranker performs relevance judgment on the relevance determination inputs to obtain the relevance ranking of candidate objects within the corresponding recall paths.

[0036] When the same candidate object appears repeatedly within the same recall path, the supply and demand matching platform retains the candidate object records with higher relevance ranking within the path, and forms a re-ranked candidate set within the path based on the retained candidate object records. Through the combined processing of "multi-path recall, hybrid retrieval, and Reranker re-ranking", each recall path can output a candidate set with ranking relationships within the path, providing a ranking basis for subsequent cross-path fusion.

[0037] S4: Merge the candidate sets for reordering within each recall path, perform reciprocal sorting and fusion based on the candidate's ranking position in the corresponding recall path, obtain preliminary fusion ranking results, filter the candidate's product subject based on the subject association information and product standardization name in the structured demand elements, input the candidate's product subject filtered and structured demand elements into the LLM model for ranking, and output the ranking results of matching enterprises.

[0038] In S4, the supply and demand matching platform first merges the candidate sets for intra-path reordering output from the enterprise profile recall path, product name recall path, product original description recall path, and product attribute recall path. During the merging process, candidate object records from different recall paths are grouped according to candidate object identifiers to obtain a cross-path candidate object set. The candidate object identifier is used to associate records from different recall paths that point to the same enterprise, product, or supply object.

[0039] After the cross-path candidate object set is formed, the supply and demand matching platform reads the ranking position of each candidate object within each recall path and marks candidates not appearing in their corresponding recall paths as non-participating. Subsequently, the supply and demand matching platform uses Reverse Ranking fusion (RRF) to perform inverse ranking fusion. Inverse ranking fusion generates a fusion ranking score based on the candidate object's ranking position in each recall path. For candidate objects appearing in multiple recall paths simultaneously, their ranking positions in all paths contribute to the fusion ranking score generation. After fusion, the supply and demand matching platform generates a preliminary fusion ranking result based on the fusion ranking score.

[0040] The preliminary sorting and screening stage is used to merge, fuse, and pre-screen the candidate results obtained from multi-path heterogeneous recall. First, for the candidate results returned from different recall paths, the Reverse Rank Fusion (RRF) method is used to uniformly fuse the multi-path candidate results. For any candidate object d, its single-path score can be expressed as: ; Here, r(d) represents the ranking position of candidate d, and k is a smoothing constant used to reduce the excessive amplification effect of high-rank differences. Finally, the RRF scores of each path are weighted and summed to obtain the preliminary fused ranking. When a candidate appears in multiple recall paths simultaneously, its RRF score will increase accordingly. In this way, multiple heterogeneous recall results can be uniformly fused without relying on the comparability of the original scores of different channels, effectively alleviating the ranking difficulties caused by inconsistent scoring standards and overlapping results.

[0041] Secondly, after fusing the results from multiple sources, the candidate products are validated for consistency based on the demand subjects, keywords, and standardized product names extracted during the demand analysis phase. Candidates whose product names significantly deviate from the demand subjects are either downgraded or eliminated entirely. This step reduces false recalls of semantically similar but entity-mismatched products. Through this pre-sorting and pre-screening process, higher-quality candidates that better align with the demand subjects and business constraints can be extracted from a large-scale candidate set, providing stable input for subsequent fine-tuning of the large-scale model.

[0042] After the initial fusion and ranking results are formed, the supply and demand matching platform performs product subject screening. Product subject screening is based on the subject association information, keywords, and standardized product names obtained during the demand analysis phase. The supply and demand matching platform reads candidate object records according to the ranking order in the initial fusion and ranking results, and extracts the product name field, product type field, and standardized product name from the candidate object records. Subsequently, the supply and demand matching platform performs a consistency comparison between the subject association information in the structured demand elements and the standardized product name to obtain subject consistency results. When the subject consistency result meets the preset consistency conditions, the corresponding candidate object is retained; when the subject consistency result does not meet the preset consistency conditions, the corresponding candidate object is downgraded, or removed when the preset consistency conditions are clearly not met and the candidate object significantly deviates from the demand subject. After product subject screening, the supply and demand matching platform obtains a set of screened candidate objects.

[0043] After screening, the candidate set enters the LLM model for fine-tuning. The supply and demand matching platform assembles the demand subjects, keywords, and attribute constraints from the structured demand elements with the company profiles, capabilities, business scope, core product information, product parameters, product attributes, geographic information, and constraint attribute matching information of each candidate in the screened candidate set, forming the candidate ranking input. The candidate ranking input can be organized through prompt templates, allowing the LLM model to make a comprehensive judgment on an object-by-object basis based on the relevance of the company's overall capabilities to the semantics of the demand, the degree of matching between the core products and the demand subjects, the degree to which product attributes meet the demand constraints, and the degree of fit between the company's geographic location and business scope.

[0044] The LLM model ranks each candidate based on the candidate ranking input, obtaining a ranking score and generating explanatory text. This explanatory text explains the match between the candidate and the fine-grained requirements of the user's needs, as well as the matching basis. The supply-demand matching platform sorts the candidates in descending order according to the ranking score, outputting the Top-N matching companies. The Top-N matching companies include the ranked candidate companies or products, and may include the matching reasons generated by the LLM model.

[0045] In the implementation of the above method, natural language demand text is first transformed into structured demand elements, which then drive four recall paths: enterprise profile, product name, original product description, and product attributes. Within each recall path, dense vector retrieval and sparse vector retrieval are first fused within the path, and then a Reranker is used to form a candidate set for re-ranking within the path. After unified fusion of multiple candidate objects using RRF, product subject filtering reduces semantically similar but mismatched core entities. Finally, an LLM model performs a comprehensive ranking within a smaller candidate set. This processing chain allows demand subjects, keywords, and attribute constraints to permeate the demand parsing, recall, fusion, filtering, and fine-tuning processes.

[0046] Please see Figure 5 This embodiment provides an AI-driven multi-path hybrid supply and demand matching and sorting system. The system is equipped with a demand processing module, a recall processing module, a sorting and filtering module, and a fine sorting output module connected along the data flow direction.

[0047] The requirement processing module receives natural language requirement text input by the user, performs requirement analysis on the text, and generates structured requirement elements representing the requirement subject, keywords, and attribute constraints. When the structured requirement elements meet preset completeness conditions, the requirement processing module passes them to the recall processing module; when the structured requirement elements do not meet the preset completeness conditions, the requirement processing module generates requirement clarification information and updates the structured requirement elements based on supplementary input. The output of the requirement processing module includes structured requirement elements and requirement clarification information. The structured requirement elements are used to construct query expressions, and the requirement clarification information is used to guide users to supplement core missing information or eliminate semantic ambiguity.

[0048] The recall processing module receives structured requirement elements and constructs multi-path query expressions based on these elements. The module generates adapted query expressions according to the enterprise profile recall path, product name recall path, product original description recall path, and product attribute recall path, and configures recall field identifiers for each query expression. The multi-path query expressions are then output to the recall paths invoked by the sorting and filtering module.

[0049] The ranking and filtering module performs hybrid recall based on multi-way query representations in the corresponding recall paths. It calls the Milvus vector database for retrieval and uses the BGE-M3 model to generate dense and sparse vector representations. In each recall path, the module obtains dense vector retrieval results based on dense vector retrieval requests and sparse vector retrieval results based on sparse vector retrieval requests, then merges these two types of results into a fused candidate result. The module further calls Reranker to re-rank the fused candidate results, forming a re-ranked candidate set within each path.

[0050] The fine-ranking output module merges the candidate sets within each recall path, using Reverse Ranking (RRF) to perform inverse ranking and fusion based on the candidate's position in the corresponding recall path, resulting in a preliminary merged ranking. The fine-ranking output module further filters the candidate objects based on the subject association information, keywords, and standardized product names in the structured requirement elements, forming a filtered candidate object set. Subsequently, the fine-ranking output module inputs the filtered candidate object set and structured requirement elements into the LLM model. The LLM model then ranks and judges the candidates based on dimensions such as capability matching, product fit, attribute satisfaction, and regional and business scope adaptation, outputting the Top-N matching enterprise results.

[0051] The specific processes, field states, processing order, judgment conditions, module collaboration, and data transfer relationships described in this embodiment are used to explain the possible implementations of the present invention and should not be limited to the specific embodiments listed. Without departing from the technical solutions described in this invention and the scope of the original disclosure, any equivalent substitutions, equivalent modifications, equivalent combinations, order adjustments, corresponding module replacements, equivalent changes in field names, equivalent transfer of execution entities, or equivalent changes in carrier forms that can be conceived by those skilled in the art should fall within the scope of protection of this patent; however, they should not be extended to unclaimed topics, and equivalent changes should not replace, delete, or weaken the method features, data relationships, temporal relationships, judgment relationships, or module collaboration relationships necessary to achieve the technical effect.

Claims

1. An AI-driven multi-path hybrid supply and demand matching and sorting method, characterized in that, The method includes: The system receives natural language requirement text input by the user, performs requirement analysis on the natural language requirement text, and generates structured requirement elements that represent the requirement subject, retrieval clues and constraint attributes. When the structured requirement elements meet the preset completeness conditions, the system proceeds to query expression construction. When the structured requirement elements do not meet the preset completeness conditions, the system generates requirement clarification information and updates the structured requirement elements based on supplementary input. Based on the structured demand elements, a multi-way query expression is constructed according to the supply-side profile field, product entity field, and constraint attribute field; Based on the multi-way query expression, hybrid recall is performed in the corresponding recall path. In each recall path, the dense vector retrieval results and the sparse vector retrieval results are fused, and the fused candidate results are reordered to form a reordered candidate set within the path. The candidate sets for reordering within each recall path are merged, and the candidate objects are merged by reversing their order position in the corresponding recall path to obtain a preliminary merged ordering result. The candidate objects are then screened for product subjects based on the subject association information and product standardized name in the structured demand elements. The candidate objects after product subject screening are then input into the LLM model for sorting with the structured demand elements, and the matching enterprise sorting result is output.

2. The AI-driven multi-path hybrid supply and demand matching and sorting method according to claim 1, characterized in that, The process of demand analysis includes: The natural language requirement text is segmented into text fragments, and candidate elements are extracted according to requirement object terms, retrieval expression terms, and constraint description terms. Write the candidate elements into the demand subject item, retrieval clue item, and constraint attribute item, and mark the candidate elements with missing fields, field conflicts, or inability to be mapped to the recall path as elements to be clarified. The requirement clarification information is generated based on the elements to be clarified, the structured requirement elements are updated after receiving the supplementary input, and the updated structured requirement elements are used as the data basis for determining whether the preset integrity conditions are met.

3. The AI-driven multi-path hybrid supply and demand matching and sorting method according to claim 2, characterized in that, The process of determining whether the structured requirement elements meet the preset integrity conditions includes: Based on the field mapping relationship corresponding to the recall path, the mapping verification is performed on the demand subject item, the retrieval clue item, and the constraint attribute item. When the demand subject item, the search clue item, and the constraint attribute item can all be mapped to at least one recall path and there are no field conflicts, it is determined that the structured demand element satisfies the preset completeness condition; When any feature item cannot be mapped to the recall path or there is a field conflict, the corresponding feature item will be identified as the feature to be clarified.

4. The AI-driven multi-path hybrid supply and demand matching and sorting method according to claim 1, characterized in that, The process of constructing the multi-way query expression includes: Based on the structured requirement elements, establish a field mapping relationship between element items and recall fields; According to the field mapping relationship, the element items representing supply capacity are mapped to the supply-side profile field, the element items representing product entities are mapped to the product entity field, and the element items representing business constraints are mapped to the constraint attribute field. Based on the field mapping results, generate the multi-way query expressions with recall field identifiers respectively, and assign the multi-way query expressions to the recall paths corresponding to the recall field identifiers.

5. The AI-driven multi-path hybrid supply and demand matching and sorting method according to claim 4, characterized in that, The process of performing the hybrid recall includes: In each recall path, a query expression corresponding to the recall path is selected from the multi-way query expressions, and dense vector retrieval requests and sparse vector retrieval requests are generated respectively. Dense vector retrieval results are obtained based on the dense vector retrieval request, and sparse vector retrieval results are obtained based on the sparse vector retrieval request. According to the fusion weight corresponding to the recall path, the dense vector retrieval result and the sparse vector retrieval result are merged into a fused candidate result, and the fused candidate result is reordered.

6. The AI-driven multi-path hybrid supply and demand matching and sorting method according to claim 5, characterized in that, The process of forming the candidate set for reordering within the path includes: Extract the candidate object text corresponding to each candidate object in the fused candidate results, and combine the natural language requirement text and the candidate object text to form the correlation determination input; By calculating the matching degree of the correlation determination input, the correlation ranking of each candidate object in the corresponding recall path is obtained. When the same candidate object appears repeatedly in the same recall path, the candidate object record with the higher correlation ranking in the path is retained, and a reordered candidate set in the path is formed based on the retained candidate object record.

7. The AI-driven multi-path hybrid supply and demand matching and sorting method according to claim 6, characterized in that, The process of performing inverse sorting and fusion on the reordered candidate set output by each recall path includes: The candidate object records in different recall paths are merged according to the candidate object identifier to obtain a cross-path candidate object set; Read the sorting position of each candidate object in the cross-path candidate object set within each recall path, and mark the candidate objects that do not appear in the corresponding recall path as non-participating; A fusion ranking score is generated based on the number of recall paths in which the candidate is in the participation state, the ranking position within the path, and the recall path weight, and the preliminary fusion ranking result is formed according to the fusion ranking score.

8. The AI-driven multi-path hybrid supply and demand matching and sorting method according to claim 7, characterized in that, The process of selecting product entities based on the preliminary fusion and ranking results includes: Read the candidate object records according to the sorting order in the preliminary fusion sorting results, and extract the product name field and product type field from the candidate object records; The consistency result of the subject association information in the structured requirement elements is obtained by comparing it with the standardized product name. When the subject consistency result meets the preset consistency condition, the corresponding candidate object is retained. When the subject consistency result does not meet the preset consistency condition, the corresponding candidate object is downgraded, and a set of filtered candidate objects is formed based on the retained or downgraded candidate objects.

9. The AI-driven multi-path hybrid supply and demand matching and sorting method according to claim 8, characterized in that, The process of inputting the filtered candidate object set and the structured requirement elements into the LLM model for sorting includes: The demand subject, search clues, and constraint attributes in the structured demand elements are combined with the supply-side profile information, product entity information, and constraint attribute matching information of each candidate object in the filtered candidate object set to form a candidate ranking input; The LLM model is used to sort the candidate sorting inputs object by object to obtain the sorting score corresponding to each candidate object. The candidate objects are sorted in descending order according to the ranking score, and the sorted candidate objects are determined as the matching enterprise ranking result.

10. An AI-driven multi-path hybrid supply and demand matching and sorting system, the system being used to implement the AI-driven multi-path hybrid supply and demand matching and sorting method as described in any one of claims 1-9, characterized in that, The system includes a demand processing module, a recall processing module, a sorting and filtering module, and a fine-sorting output module connected along the data flow direction. The requirement processing module receives natural language requirement text input by the user, performs requirement analysis on the natural language requirement text, generates structured requirement elements that represent the requirement subject, search clues and constraint attributes, and enters the query expression construction when the structured requirement elements meet the preset completeness conditions. When the structured requirement elements do not meet the preset completeness conditions, requirement clarification information is generated and the structured requirement elements are updated according to supplementary input. The recall processing module constructs a multi-way query expression based on the structured demand elements, according to the supply-side profile field, product entity field, and constraint attribute field. The sorting and filtering module performs hybrid recall in the corresponding recall path based on the multi-way query expression. In each recall path, the dense vector retrieval results and the sparse vector retrieval results are fused, and the fused candidate results are re-sorted to form a re-sorted candidate set within the path. The fine-ranking output module merges the re-ranking candidate sets output by each recall path, performs reciprocal sorting and fusion based on the ranking position of the candidate objects in the corresponding recall path, obtains a preliminary fusion ranking result, filters the candidate objects for product subjects based on the subject association information and product standardized names in the structured demand elements, inputs the candidate objects after product subject filtering and the structured demand elements into the LLM model for ranking, and outputs the matching enterprise ranking result.

Images